Breaking the Bottleneck | Issue 82
[6/16/2024] Tariffs Impact, Robot Revelations, Waymo Robotaxi Factory, the IEA EV Report, Nominal Announcement & More!
Breaking the Bottleneck is a weekly manufacturing technology newsletter with perspectives, interviews, news, funding announcements, manufacturing market maps, 2025 predictions, and more!
💥 If you are building, operating, or investing in manufacturing, supply chain, or robots, please reach out - adityabreakingthebottleneck@gmail.com I’d love to chat!
🏭 If you were forwarded this and found it interesting, please sign up!
Interview & Startup Series 🎙️💬
Alongside the usual newsletter, I’m excited to launch a weekly interview series, “Friday 5,” starting in July. The series focused on navigating the industrial software landscape, featuring industrial operators, founders, tech providers, and more. If you, “Friday 5, u are similarly enthusiastic and know people that I should interview, I’d love to hear from you and connect!
Content I Enjoyed Last Week 🗞️🔬 📚
News:
Inside The Waymo Factory Building A Robotaxi Future [Forbes]
Waymo, Alphabet's autonomous driving unit, has established Mesa, Arizona, as the production hub for its expanding fleet of robotaxis. At a 239,000-square-foot facility east of Phoenix’s Sky Harbor Airport, Waymo is building and equipping Jaguar I-PACE electric SUVs with its proprietary self-driving technology, including in-house-designed lidar, radar, cameras, and computing systems. Managed by automotive veteran Kent Yiu and operated in partnership with Magna, the factory opened in October 2023 and is expected to scale up to produce tens of thousands of robotaxis annually. Though smaller than traditional car plants, the facility supports a much higher utilization rate: each of Waymo’s ~1,500 vehicles provides an average of 24 rides daily, totaling over 250,000 paid rides weekly. As production scales to 10,000 vehicles, the company could facilitate over 1.5 million rides per week, potentially generating $2 billion in annual revenue, up from an estimated $100 million last year. Waymo's expansion strategy is aggressive, with recent launches in Los Angeles, San Francisco, and Austin, and planned service in Atlanta, Miami, and Washington, D.C. The company has also begun testing in cities such as Nashville and Tokyo. Alongside the Jaguars, Waymo will add two new vehicle types to its fleet, a Zeekr van from Geely and a Hyundai Ioniq 5. While Tesla claims it can undercut Waymo’s costs using simpler sensor systems, Waymo insists its more advanced tech ensures superior safety and reliability.
How Much Energy the AI Industry Uses [MIT Tech Review]
MIT conducted an excellent study interviewing two dozen experts, measuring AI’s energy demand, evaluating different AI models and prompts, poring over hundreds of pages of projections and reports, and questioning top AI model makers about their plans. The study reveals that AI’s rapidly growing energy demands are vastly underreported and misunderstood due to a lack of transparency from leading tech companies. Inference, when users interact with models like ChatGPT, now consumes 80–90% of AI’s computing energy, and the scale is staggering: OpenAI’s GPT-4 reportedly consumed 50 GWh to train, and inference is ongoing across ~3,000 U.S. data centers. AI models are deployed on energy-hungry GPUs, such as Nvidia’s H100, and the size and complexity of the model significantly impact energy consumption. For instance, Meta’s Llama 3.1 405B uses over 6,700 joules per text query, hundreds of times more than smaller models. Image and video generation consume even more, with video creation requiring millions of joules per clip. Multiply this by billions of queries (ChatGPT alone receives 1 billion messages and 78 million images daily), and AI's electricity use becomes massive, potentially exceeding 326 terawatt-hours annually by 2028, or 12% of the U.S.'s electricity consumption. This energy surge is reshaping global infrastructure: Microsoft and Meta are exploring nuclear energy, OpenAI’s Stargate initiative targets 10 multi-gigawatt data centers (each exceeding New Hampshire’s total power demand), and global buildouts are booming. Yet, these data centers often operate on carbon-intensive grids, which are 48% more polluting than the U.S. average, raising serious environmental concerns. Governments, researchers, and utilities remain in the dark, as tech firms treat energy usage as trade secrets. Consequently, the public may ultimately bear the cost. Utility deals could shift costs to residential ratepayers, with estimated monthly hikes of $37.50 in states like Virginia.
Scaling Helix: New State of the Art in Humanoid Logistics [Figure]
In just three months, Helix, Figure’s logistics robotic system, has achieved near-human levels of dexterity and speed in package handling through a combination of data scaling and architectural advances. Initially limited to rigid boxes, Helix now reliably processes deformable packages, such as poly bags and padded envelopes, adapting its grasp and manipulation strategies in real-time. The average handling time per item has dropped from ~5.0 seconds to 4.05 seconds (a 20% improvement), and barcode orientation accuracy has increased from ~70% to 95%, aided by behaviors learned from demonstrations, such as smoothing wrinkled labels before scanning. These improvements were driven by enhanced visuo-motor architecture, including (1) vision memory for temporal context and multi-step planning, (2) state history for continuity and reactive control, and (3) force feedback for touch-aware manipulation. The training data was also expanded from 10 to 60 hours of demonstrations, resulting in a 58% improvement in throughput and a 6.2% increase in barcode success, from 88.2% to 94.4%. An ablation study confirmed that each architectural component contributed meaningfully to performance: vision memory eliminated redundant motions, state history enabled faster corrections, and force feedback improved grip precision and motion control. Notably, these enhancements allowed the same model to learn new context-driven behaviors, such as human–robot handoffs, from just a few additional demonstrations without reprogramming. The result is a flexible, general-purpose policy capable of robust, high-speed manipulation in dynamic environments. Helix’s gains highlight the power of scaling both data and model architecture simultaneously, moving closer to real-world autonomous logistics operations.
Factory Work is Overrated. Here are the Jobs of the Future [Economist]
Despite political nostalgia and recent tariff-driven efforts to revive American manufacturing, the reality is that factory jobs, as they existed in the 20th century, are not coming back. While the U.S. still produces more goods than ever (over twice the real output of the early 1980s), automation and productivity gains have sharply reduced the need for human labor on factory floors. Today, fewer than 4% of American workers are in production roles, down from nearly 25% in the 1970s. Even within manufacturing, many jobs are now in support or professional services. Globally, even industrial powerhouses like China, Germany, and South Korea have experienced steep declines in manufacturing employment, driven by automation and a shift in consumption toward services. Efforts to "reshore" manufacturing, such as imposing tariffs to close the $1.2T trade deficit, would come at a steep cost ($200,000 per job saved) and produce only modest employment gains. At the same time, modern factory work lacks the union protections, wage premiums, and accessibility it once offered. Instead, the closest modern analogs to those middle-class, non-college factory jobs are in skilled trades, such as electricians, HVAC technicians, mechanics, and emergency services. These roles offer solid pay (median approximately $25/hour), higher unionization rates, and strong demand, but lack the geographic concentration and local economic impact typically found in traditional manufacturing towns. While these jobs support individual workers, they don’t anchor regional economies the way steel or auto plants once did. Moreover, future job growth for non-degreed workers is expected to come mainly from lower-paid sectors, such as healthcare support and personal care, rather than manufacturing. As Harvard’s Dani Rodrik suggests, the challenge now is to raise productivity and wages in growing sectors rather than revive fading ones. Much like farming in Jefferson’s time, the symbolic and economic centrality of factory work has waned, leaving working-class prosperity to be redefined elsewhere.
Introducing Tulip MCP Server [Tulip]
Tulip has launched its Model Context Protocol (MCP) server. This open-source integration layer connects large language models (LLMs) with Tulip’s manufacturing platform, enabling context-aware, real-time AI interactions within production environments. The MCP standard acts as middleware, enabling AI to interact securely with external systems by exposing Tulip resources (e.g., stations, machines, tables) as tools that AI agents can access through governed APIs. This eliminates the need for custom integrations, allowing AI to autonomously retrieve metrics, trigger actions, or update data in Tulip via natural language prompts. Tulip’s implementation simplifies operational automation, reduces manual effort, and allows AI to become “Tulip-aware” for intelligent execution of manufacturing workflows. Built for both experimentation and production, the Tulip MCP server is easy to deploy (via a four-step setup on GitHub) and entirely governed by Tulip’s API permission scopes. It enables controlled access to Tulip functionality, ensuring that AI actions remain within defined security boundaries. Practical use cases include real-time production updates, automated quality insights, and rapid provisioning of workstations. For instance, AI can analyze defect rates or automatically configure new stations based on input data. Tulip encourages users to start with small tasks, such as listing tables or adding records, and provides comprehensive documentation to support seamless integration.
Tariffs Aimed at Reviving Manufacturing Are Doing the Opposite [Bloomberg]
President Trump’s aggressive tariff policy, intended to reinvigorate U.S. manufacturing and the Rust Belt, is causing disruption and uncertainty across America’s industrial heartland. Companies such as Rockwell Automation, Snap-on, and TCCI Manufacturing are delaying key investments due to volatile tariff conditions and concerns about economic instability. Manufacturing payrolls declined by 8,000 last month, and U.S. factory activity has contracted for three consecutive months, with executives attributing the decline to supply chain disruptions that rival those of the pandemic, mainly due to tariffs. The Midwest, which lost nearly 2 million manufacturing jobs between 1998 and 2010, has shown signs of recovery in recent years, driven by reshoring efforts and incentives for electric vehicles (EVs) and clean energy. However, the new tariffs are raising costs, chilling investment, and threatening progress. Rockwell Automation reports delayed customer projects, while Snap-on says even its supportive customers are growing anxious. Autodesk’s CEO warns that constant policy shifts are stalling long-term investments. Despite Trump touting tariffs as beneficial, such as increasing duties on steel to 50% during a visit to Pittsburgh, private construction spending on manufacturing has plateaued. Faribault Mill in Minnesota sees increased domestic interest but remains worried about broader economic headwinds. TCCI’s $45 million factory upgrade in Illinois was overshadowed by steep and fluctuating tariffs on Chinese equipment, delaying critical purchases. Illinois Governor JB Pritzker, speaking at the plant’s reopening, criticized the unpredictable tariff regime as detrimental to strategic manufacturing growth. While some Southern states are attracting global manufacturers with lower labor costs, the Midwest’s rebound remains fragile, threatened by the very trade policies meant to protect it. The broader concern is that tariff volatility could derail industrial momentum just as companies were regaining confidence, putting jobs, innovation, and domestic manufacturing expansion at risk.
The Impact of US Tariffs of 50% on Steel and Aluminum [BCG]
The recent increase in US tariffs on steel and aluminum, from 25% to 50%, is already reshaping global metals markets and supply chains. Between the initial announcement in February and late May, US price premiums rose sharply, with steel prices widening by 77% over EU levels and aluminum by 139%. As a result, companies are investing further in the US. Emirates Global Aluminum announced a new US facility, while Hyundai Steel and Posco are investing in a steel plant in Louisiana. US customers are also shifting strategies, such as increasing the use of domestic raw steel or reconsidering aluminum packaging. If demand remains steady, prices are likely to continue rising. Over the medium to long term, the tariff hike could push some foreign products, such as hot-rolled coil steel from the EU, out of the US market entirely, accelerating domestic production and reshaping downstream metal fabrication. However, niche, high-value products like tinplate or tool steel may retain competitiveness in exports to the US. The increase highlights growing complexity in US trade policy, particularly around Section 232 measures. Tariffs are no longer uniform; countries like the UK have negotiated exemptions, and the combination of various trade instruments makes compliance and strategic planning more challenging. According to BCG, metals customers should establish a “tariff command center” to simulate trade scenarios, manage compliance, and reshape their supply chains. US producers might shift their product mix toward value-added offerings and explore operational efficiencies or capacity expansions. Meanwhile, US customers may need to redesign products or localize production to mitigate the impacts of tariffs. Non-US exporters must reevaluate which market segments remain viable or consider US-based manufacturing to maintain access.
Other News:
Siemens and Nvidia Expand Partnership to Accelerate AI Capabilities
Honeywell Rolls Out AI-Powered Cyber Solutions [Honeywell]
World’s First Industrial AI Cloud to Advance European Manufacturing [NVIDIA]
AI in Process Manufacturing [Microsoft]
Blog / Research:
Real-Time Action Chunking with Large Models [Physical Intelligence]
Vision-Language-Action (VLA) models, while promising for generalization and reasoning in robotics, face a fundamental challenge in real-time operation: the world continues to evolve while the model is computing. Delays between observations and actions can lead to errors, particularly on mobile edge devices that rely on cloud-based inference for processing. Traditional approaches, such as action chunking (executing 50 pre-planned actions per inference), mitigate some latency but still suffer from discontinuities between chunks that can lead to hazardous robot behavior. Prior attempts to smooth transitions, like temporal ensembling, often fail catastrophically. To avoid this, earlier models, such as π0 and π0.5, ran synchronously, introducing visible pauses that degraded task performance. To address this, researchers introduced Real-Time Chunking (RTC), a method that maintains consistent trajectories between overlapping action chunks using a technique inspired by image inpainting. RTC freezes the overlapping initial actions of a new chunk and conditions the rest on both new observations and partial attention to prior actions, enabling smooth transitions without retraining existing models. It eliminates pauses between chunks, maintains high performance even under artificial inference delays of over 200ms, and outperforms synchronous methods and prior smoothing baselines. A new throughput metric (substeps per minute) reveals that RTC maintains task performance, whereas others degrade sharply with increasing latency. Additional evaluation using controller steps (inference time excluded) reveals that RTC not only completes tasks faster but also makes fewer mistakes earlier in execution. These findings underscore that as robot foundation models scale and require more inference time, robust real-time execution strategies, such as RTC, will be critical. Future efforts will need to build on this by enabling hierarchical inference, multi-timescale planning, and dynamic reallocation of computational resources to bring scalable, general-purpose physical intelligence to real-world robotics.
V-JEPA 2 World Model and New Benchmarks for Physical Reasoning [Meta]
Meta's V-JEPA 2 (Video Joint Embedding Predictive Architecture 2) is a 1.2 B-parameter self-supervised video-based world model designed to enable AI agents to understand, predict, and plan actions in the physical world. Built on Meta’s JEPA architecture, V-JEPA 2 advances zero-shot robot control, enabling tasks like object manipulation in unfamiliar settings without requiring fine-tuning on specific environments. The model is trained in two phases: actionless pre-training on over 1 million hours of diverse video and images to build a semantic understanding of physical interactions, followed by action-conditioned training using only 62 hours of robot data to teach control dynamics. This approach enables V-JEPA 2 to perform short-horizon tasks via model predictive control using visual goal states, and longer tasks by sequencing subgoals, achieving success rates of 65%–80 % for pick-and-place tasks in unseen environments. It also achieves state-of-the-art results in video understanding tasks, such as action anticipation and video Q&A, when combined with language models. To benchmark physical reasoning capabilities, Meta introduced three new open-source datasets: IntPhys 2, MVPBench, and CausalVQA. IntPhys 2 evaluates intuitive physics by requiring models to detect implausible events in paired videos, revealing that even top models perform near-chance levels compared to human near-perfect accuracy. MVPBench tests video-language models with minimal-change question pairs designed to avoid shortcut learning, while CausalVQA probes causal reasoning and counterfactual understanding. These benchmarks highlight gaps in predictive and planning abilities that are critical for Advanced Machine Intelligence (AMI). Looking ahead, Meta plans to develop hierarchical and multimodal JEPA models that can learn across various timescales and modalities, including vision, audio, and touch, to better emulate human-like world modeling and decision-making. All models, benchmarks, and leaderboards are publicly released via GitHub and Hugging Face to foster community progress.
More Robotics News:
Galbot Introduces OpenWBT, the open-source, whole-body VR teleoperation system designed for humanoids across various robot types & virtual/physical worlds.
1X Launches Redwood, a vision-language transformer designed for the humanoid form factor, capable of performing end-to-end mobile manipulation tasks such as retrieving objects for users, opening doors, and navigating around the home.
How Data Travels in F1 [Idee Fixe]
Podcast/Video:
I Tried to Make Something In America
Finance & Transactions 💵
Venture Capital:
Gecko Robotics - A company using fixed sensors and robots that climb, crawl, swim, and fly to build first-order data layers on the physical world.
$125 million [Series D] - Led by Cox Enterprises and joined by USIT, XN, Founders Fund, YC, and Friends & Family Capital.
Nominal - A company building a Rust-based platform for hardware testing, verification, and validation.
$75 million [Series B] - Led by Sequoia, joined by Lightspeed, Lux Capital, General Catalyst, and Founders Fund.
Applied Computing - A company helping oil, gas, and petrochemical plants optimize yields, reduce emissions, and prevent failures using real-time ops data.
$12.2 million [Seed] - Led by Stride VC
ZeroRISC - A company building silicon supply chain integrity solutions.
$10 million [Seed] - Led by Fontinalis Partners
Bolo AI - A company building a complete AI solution specially designed for the energy industry to make your operations faster, safer, and more efficient
$8.1 million [Seed] - Led by True Ventures and joined by Benchstrength, Accomplice, J Ventures, and Beat Ventures
Sunrise Robotics - A Slovenian company building modular industrial robotics and AI models that make them simple to deploy in different manufacturing environments.
$8.5 million [Seed] - Led by Plural and joined by Tapestry, Seedcamp, Tiny.vc, and Prototype Capital.
Saeki - A company building autonomous factories that combine large-format 3D printing, CNC machining, and in-process inspection into autonomous cells
$6.7 million [Seed] - Led by Lightbird and joined by Founderful, 2100VC, and Danobat.
Planned Downtime 🧑🔧
The Semiconductor Century with Chris Miller
Great article. Curious about the breakdown for 4% in : "fewer than 4% of American workers are in production roles" in terms of role and industry (cannot read the original content).